Erdosproblems.com Succumbs to the AI Onslaught

TEXT START: I want to start by reiterating my comment in Thomas Bloom's initial "what should I change" post; he has done a wonderful service in creating this site and making these problems, and their history, broadly accessible and organized.

The Dissection

The text defends open visibility as mathematical infrastructure: public proof claims, historical status, and low-friction comments create discovery and collaboration. Its real target is not moderation itself but the conversion of a living research commons into a filtered archive.

The author correctly identifies a genuine loss: aggressive filtering can destroy weak-tie collaboration, serendipitous discovery, and the informal reputation signals that make obscure contributors visible. The argument becomes obsolete, however, when it treats AI-generated mathematical output as merely more of the same human slop.

The Core Fallacy

The central error is assuming that visibility remains productive when the cost of generating claims collapses. In the AI age, an open forum is not simply a larger pool of ideas. It becomes a high-volume attack surface against human attention. The scarce resource is no longer proof production; it is trusted verification and prioritization.

The author treats reduced visibility as the primary bottleneck. Under AI-driven abundance, the bottleneck is signal extraction. Unfiltered access does not automatically preserve progress. It can bury progress beneath claims that are cheap to generate and expensive to check.

The argument also assumes that collaboration requires the old comment architecture. That is a historical contingency, not a mathematical law. If the site cannot distinguish useful conjectures from machine-produced debris, its openness becomes a coordination failure rather than a commons.

Hidden Assumptions

• Human attention can scale with the volume of AI-assisted claims.
• Most marginal proof submissions have enough expected value to justify their verification cost.
• Readers can reliably identify useful ideas before the noise overwhelms them.
• Zenodo's invisibility is the main problem, rather than the absence of trusted indexing, reputation, and verification layers.
• Previous collaboration rates will persist after the economics of proof generation change.
• “Slop” and valuable unconventional work can be separated cheaply without imposing filters.
• The site’s historical openness can survive without becoming a dumping ground for automated output.

These assumptions confuse the conditions that made the old system productive with conditions that will remain true after cognitive production is automated.

Social Function

Classification: partial truth and transition-management resistance.

The text records a real casualty of the transition: centralized filtering can kill discovery and exclude outsiders. But it also functions as a defense of the pre-AI prestige and attention regime, where being visible was relatively cheap to administer because human output was limited. It asks the old commons to absorb industrial-scale synthetic production without changing its control system.

That is not preservation. It is an attempt to keep the corpse’s public square open while the cost of filling it with noise approaches zero.

The Verdict

The complaint is strongest as a warning about collateral damage: indiscriminate suppression can erase valuable human work and destroy collaboration channels. It fails as a systemic diagnosis because it treats openness as inherently pro-progress after AI has made production abundant and verification scarce.

The old site model is not necessarily dead, but its unfiltered form is terminal. Mathematical progress now requires trusted triage, provenance, machine-checkable verification, and reputation systems capable of separating genuine leverage from automated exhaust. Without those layers, “more visible” does not mean more knowledge. It means a thicker fog.